This study investigated the accuracy of video-based observational posture analysis for the elbow and shoulder. Posture analyses were conducted by 28 ergonomists for four jobs presented on a traditional VHS format video recording. Estimates of posture from the observational-based methods were compared with values measured directly with an optical motion capture system. Ergonomists used categorical posture scales and a continuous visual analog scale to estimate the peak and most frequently occurring or average posture for each job. Use of a three-category scale resulted in misclassifications of peak and most frequently occurring elbow and shoulder posture with a probability averaging 30.1%. With the six-category posture scale this average probability of misclassification increased to 64.9%. Using a continuous visual analog scale peak shoulder elevation was the only posture for which the average error among ergonomists' estimates was significantly different from zero (p<0.05). Correlations between the estimated postures and measured postures were higher and statistically significant (p<0.05) for elbow flexion and shoulder elevation (r2 between 0.45 and 0.66) but were considerably lower and not significant (r2 between 0.03 and 0.18) for the peak and average horizontal shoulder abduction. Ergonomists' estimates of the temporal distribution of shoulder posture, indicating the duration severity of the posture, appeared to be biased such that the percentage of the cycle time in each posture category was estimated as more uniformly distributed than the measured values indicated.